High Performance Multivariate Geospatial Statistics on Manycore Systems

Modeling and inferring spatial relationships predicting missing values of environmental data are some the main tasks geospatial statisticians. These routine accomplished using multivariate models cokriging technique. The latter requires evaluation expensive Gaussian log-likelihood function, which has impeded adoption for large datasets. However, this large-scale challenge provides a fertile ground supercomputing implementations statistics community as it is paramount to scale computational capability match growth in coming from widespread use different collection technologies. In article, we develop deploy modeling inference on parallel hardware architectures. To tackle increasing complexity matrix operations massive concurrency systems, leverage low-rank approximation techniques with task-based programming schedule asynchronous dynamic runtime system. proposed framework both dense approximated computations function. It demonstrates accuracy robustness performance scalability variety computer systems. Using synthetic real datasets, shows better compared exact computation, while preserving application requirements parameter estimation prediction accuracy. We also propose novel algorithm assess after online estimation. quantifies benchmark measuring efficiency several modeling.